Every december Spotify makes you a “wrapped” playlist of the 100 songs you most frequently listened to that year giving a clear picture of your music taste that year. I’ve always found that my music taste is kind of variable with a lot of different genre’s so it would be interesting to see how my music taste has changed from 2016 till now. To compare my music taste from 2016 to 2020 i will analyze my Spotify wrapped 2016-2020 playlists using the following metrics: valence, danceability, energy, tempo, key and modality.
In addition i’d like to analyze my most frequently listened song and one favorite song or outlier from every playlist with the use of self similarity matrices, classifiers, tempo, chroma and chordograms.
My Corpus is divided into 5 groups of spotify wrapped playlists representing 2016 to 2020:
My music from 2016-2020 has generally become more and more danceable every year shifting to the right until in 2020 almost all songs have a danceability higher than 0.4. The energy distribution stays more or less the same every year.
As mentioned in the previous tab you can see the music used for sleeping in the 2019 plot at the bottom left it has a clear low score on just about every variable.
2016 Jesus of Suburbia - Green Day: a favorite because the song has lots of different sections and tempo changes. You can clearly see some of the different sections in the tempogram with tempo changes at 100, 200 and 400 seconds. However the bpm is way off starting at 300 BPM and almost reaching 400 at 400 seconds in. While the average BPM is around 147.
2017 Weightless part 1 - Marconi Union: an outlier because its sleep music and is not something i would listen to on the regular. As expected Spotify can’t really figure out the tempo as this song is just atmospheric sounds without a distinct way to measure tempo.
2018 Bohemian Rhapsody - Queen: a favorite because it has different sections and tempo changes. You can clearly see the segment change where the staccato piano comes in and the tempo changes at around 180 seconds in and when the guitars solo’s start at around 250 seconds in. Just as for every tempogram before the tempo seems way off with the song having an average bpm of 72, while the tempogram shows 2 lines one at around 100-110 bpm and one at around 300 bpm possibly indicating that spotify has trouble identifying the tempo of songs that have these kinds of tempo changes.
2019/2020 Build God, Then We’ll Talk - Panic! At The Disco: a favorite because it has different sections, tempo and time signature changes making it a very unique song. At 50 seconds in, the time signature changes to a 3/4 and then reverts back to 4/4 after a small section then another time signature change just after the 1 minute mark explains the small gap in the line. Spotify correctly finds the tempo change at 130 seconds. Once again the BPM is totally wrong with this song having an average tempo of 124.
Outliers:
2016 Jesus of Suburbia - Green Day: a favorite because the song has lots of different sections and tempo changes.
2017 Weightless part 1 - Marconi Union: an outlier because its sleep music and is not something you would listen to on the regular.
2018 Bohemian Rhapsody - Queen: a favorite because it has different sections and tempo changes.
2019/2020 Build God, Then We’ll Talk - Panic! At The Disco: a favorite because it has different sections, tempo and time signature changes making it a very unique song.
2019 was excluded due to errors.
Overall the classifiers don’t work very well for my corpus but i expected as much. Since i do listen to a lot of different music the algorithms might have trouble finding distinct features for every year the different years. However, with limited features on the random forest classifier 2016, 2017 and 2020 seem to be working a lot better, almost classifying 50% correctly. Another thing worth noting is the correlation between 2016 and 2018 for both classifiers. The classifiers keep labeling 2016 as 2018 and vice versa. The k-nearest neighbors classifier was trained on all of the features and the random forest trained only on the top half of the features from the feature importance plot.